Exploring adaptive memory layer solutions for LLM agents over long-horizon settings. Working with Prof. Lucy Wang on replicating memory evals for coding agents with ~1M context, and improving memory recall for long trajectories.

Asked myself: "How much data is enough data?". Conducted data curation experiments on 6 languages by finetuning only on low-perplexity utterances and found that using high-quality curated data often leads to faster convergence as opposed to full finetuning.

Led an end-to-end ASR project on 21 extremely low-resource languages (under 10 hours of spontaneous Mozilla Common Voice speech). Fine-tuned MMS-1b, XLS-R-1b, and Whisper-large, built a KenLM n-gram ARPA pipeline and integrated pyctcdecode beam search with shallow fusion. 4-gram LM decoding beat greedy decoding by up to 27%. Presented as an oral paper at at ACL 2026.

Implemented language modelling from scratch using Python and PyTorch. Most interesting finding was how much tokenization can be made more efficient through proper use of data structures and multithreading. See blog and code!

This paper presents a comprehensive framework that standardizes the dataset creation pipeline across seven critical checkpoints by identifying systemic challenges in hate speech dataset creation.

I contributed to a shared task focused on identifying homophobic and transphobic content in YouTube comments by implementing basic classifiers using multilingual pre-trained language models to analyze English, Tamil, and code-mixed datasets.